Samsung SAIL Montréal

Summary: Samsung AI Lab in Montréal (SAIL = Samsung AI Lab). Produced HRM (Hierarchical Reasoning Model) and TRM (Tiny Recursive Model) — recursive reasoning architectures achieving SOTA on ARC-AGI with tiny parameter counts (7M-27M) via deep supervision + ACT. Key researcher: Alexia Jolicoeur-Martineau.


Details

For Labs

  • Type: Industrial Research Lab (Samsung)
  • Location: Montréal, Canada
  • Key Research: Recursive reasoning, deep supervision, ARC-AGI, ACT, small-model reasoning
  • Key Papers: HRM (2025), TRM (2025)
  • Lead Researcher: Alexia Jolicoeur-Martineau
  • Website: samsonailabs.com (Samsung AI Labs)

Significance

Research alignment: Your interest in recursive reasoning, latent space reasoning, ACT, deep supervision directly matches their output.

HRM/TRM Innovation Your Interest Match
Deep supervision (primary driver, 19%→39%) Core mechanism for your recursive models
ACT with Q-learning halting Latent space reasoning + adaptive compute
1-step gradient (IFT + Neumann) Efficient training for deep recursion
7M-27M params beating LLMs on ARC Small-model reasoning — very relevant

PhD relevance:

  • Academic + industry hybrid (Samsung + Mila/Université de Montréal ecosystem)
  • Publishes at top venues (HRM/ICLR, TRM/arXiv)
  • Small team → potential for direct mentorship
  • Montréal = top AI hub (Mila, McGill, UdeM, Google DeepMind, Meta FAIR, Microsoft)

Target lab assessment: Tier 1. Perfect match for recursive reasoning / ACT / deep supervision research. Alexia Jolicoeur-Martineau is emerging star.


Key Papers

Paper Year Innovation Your Wiki
HRM: Hierarchical Reasoning Model 2025 H/L hierarchy, deep supervision, ACT paper-hrm
TRM: Tiny Recursive Model 2025 Single tiny net, deep supervision is key paper-trm

Key People

Person Role Your Wiki
Alexia Jolicoeur-Martineau Lead Researcher (HRM, TRM) alexia-jolicoeur-martineau

Research Directions (from papers)

  1. Deep supervision as primary driver — TRM ablation shows hierarchy adds little; deep supervision is key
  2. ACT stability via Post-Norm + AdamW — avoids replay buffers/target networks
  3. Inference-time scaling via ACT — train M_max=8, test M_max=16 zero-shot
  4. Flat recursion > hierarchy — TRM beats HRM with 4× fewer params
  5. Small models for reasoning — 7M params competitive with 7B LLMs on ARC

Contact Strategy

Timing: Pre-application (Fall 2025) — reach out Summer 2025 with paper/idea Angle: Your DSA-ViT work + interest in recursive vision reasoning + ACT for visual tasks Hook: TRM removes hierarchy but keeps deep supervision → apply to ViT? Visual ARC? Video reasoning?


Related Wiki Pages


Sources

  • arxiv-2510.04871: TRM paper (affiliation)
  • arxiv-2506.21734: HRM paper (affiliation)